Moderation Challenges in Voice-based Online Communities on Discord - arXiv.org

Moderation Challenges in Voice-based Online Communities
 on Discord
 JIALUN “AARON” JIANG, University of Colorado Boulder, USA
arXiv:2101.05258v1 [cs.HC] 13 Jan 2021

 CHARLES KIENE, University of Washington, USA
 SKYLER MIDDLER, University of Colorado Boulder, USA
 JED R. BRUBAKER, University of Colorado Boulder, USA
 CASEY FIESLER, University of Colorado Boulder, USA
 Online community moderators are on the front lines of combating problems like hate speech and harassment,
 but new modes of interaction can introduce unexpected challenges. In this paper, we consider moderation
 practices and challenges in the context of real-time, voice-based communication through 25 in-depth interviews
 with moderators on Discord. Our findings suggest that the affordances of voice-based online communities
 change what it means to moderate content and interactions. Not only are there new ways to break rules that
 moderators of text-based communities find unfamiliar, such as disruptive noise and voice raiding, but acquiring
 evidence of rule-breaking behaviors is also more difficult due to the ephemerality of real-time voice. While
 moderators have developed new moderation strategies, these strategies are limited and often based on hearsay
 and first impressions, resulting in problems ranging from unsuccessful moderation to false accusations. Based
 on these findings, we discuss how voice communication complicates current understandings and assumptions
 about moderation, and outline ways that platform designers and administrators can design technology to
 facilitate moderation.
 CCS Concepts: • Human-centered computing → Empirical studies in collaborative and social com-
 puting; Social networks; Social networking sites;
 Additional Key Words and Phrases: moderation; voice; online communities; gaming communities; ephemerality;
 Discord
 ACM Reference Format:
 Jialun “Aaron” Jiang, Charles Kiene, Skyler Middler, Jed R. Brubaker, and Casey Fiesler. 2019. Moderation
 Challenges in Voice-based Online Communities on Discord. Proc. ACM Hum.-Comput. Interact. 3, CSCW,
 Article 55 (November 2019), 23 pages. https://doi.org/10.1145/3359157

 1 Introduction
 Online communities face malicious behaviors like hate speech [12] and harassment [8], and many
 community moderators volunteer their time to combat these problems on a daily basis. People in
 online communities usually have a good sense of how moderation works: Someone posts something
 Authors’ addresses: Jialun “Aaron” Jiang, University of Colorado Boulder, Department of Information Science, ENVD 201,
 1060 18th St. Boulder, CO, 80309, USA, aaron.jiang@colorado.edu; Charles Kiene, University of Washington, Department of
 Communication, Seattle, WA, 98195, USA, ckiene@uw.edu; Skyler Middler, University of Colorado Boulder, Department of
 Information Science, ENVD 201, 1060 18th St. Boulder, CO, 80309, USA, skyler.middler@colorado.edu; Jed R. Brubaker,
 University of Colorado Boulder, Department of Information Science, ENVD 201, 1060 18th St. Boulder, CO, 80309, USA,
 jed.brubaker@colorado.edu; Casey Fiesler, University of Colorado Boulder, Department of Information Science, ENVD 201,
 1060 18th St. Boulder, CO, 80309, USA, casey.fiesler@colorado.edu.

 Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
 provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the 55
 full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored.
 Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires
 prior specific permission and/or a fee. Request permissions from permissions@acm.org.
 © 2019 Copyright held by the owner/author(s). Publication rights licensed to ACM.
 2573-0142/2019/11-ART55 $15.00
 https://doi.org/10.1145/3359157

 Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

55:2 J. A. Jiang et al.

inappropriate; either it is seen and removed by a human moderator, or instantly removed by an
automated moderator. This process is straightforward and easy to understand, and has long been
part of people’s mental models. But how does a moderator, whether human or automated, moderate
inappropriate speech when it is spoken, in a real-time voice chat rather than text that can be erased?
 Adoption of new technology and new communication media introduces new ways for community
members to engage and communicate, but it also introduces new norms to the communities [42],
and consequently changes what it means to moderate content and interactions. Consider real-time
voice chat. This is not a new technology, but only recently with the increasing popularity of Discord
and other voice-based online communities, has become more relevant for the everyday process
of content moderation. In traditional text-based communities, moderation work mostly involves
moderators locating the problematic content, and then removing it and sometimes also punishing
the poster. This is a process that many people would take for granted, but how does this process
work in the context of real-time voice, a type of content that lasts for a short time without a
persistent record? The moderation of ephemeral content raises a number of questions: How do
moderators locate the content? How do moderators remove the content? How do moderators know
who the speaker is? How do moderators know whether the rule breaking happened at all?
 Voice has exposed new problems for moderation, particularly due to the absence of a persistent,
written record as is common in other major large-scale moderated spaces. With voice and its strict
ephemerality making current content management options impossible, voice moderators have to
face these questions every day, and develop their own tactics and workarounds. Though Discord is
designed for voice communication, Discord users have appropriated the platform technology to
also play other types of audio (e.g., music or even noise) have become a part of user interaction as
well, and we will describe ways that this also plays into types of rule-breaking. Furthermore, these
issues do not impact voice alone—answers to these questions will not only provide insights toward
how voice moderation works, but also provide insights for a range of emerging types of technology
where interactions and content are ephemeral, such as immersive online games (e.g., Fortnite) and
virtual reality (VR). These insights will also inform design and policy for communities that adopt
new technology in order to help mitigate potential moderation problems.
 In this work, we investigate the work of moderators and the challenges they experience in real-
time voice-based online communities on Discord, a popular voice-over-IP (VoIP) platform. Through
an analysis of 25 interviews with volunteer moderators of Discord communities, we first describe
new types of rules unique to voice and audio-based communities and new ways to break them.
Then, we describe how moderators struggled to deal with these problems. Moderators tried to give
warnings first but sometimes had to take actions based on hearsay and first impressions. To avoid
making elaborate rules for every situation, moderators instead simply stated that they had highest
authority. We then detail how these problems point to moderators’ shared struggle—acquiring
evidence of rule breaking, and how moderators’ evidence gathering strategies could fail in different
scenarios.
 Through the lens of Grimmelman’s taxonomy of community moderation [22] that focuses
on techniques and tools, we argue that voice precludes moderators from using the tools that
are commonplace in text-based communities, and fundamentally changes current assumptions
and understandings about moderation. From here, we provide recommendations for designing
moderator tools for voice communities that automate much of moderators’ work but still ultimately
put humans in charge, and finally argue that platform administrators and moderators must consider
the limitations imposed by the technological infrastructure before importing existing rules and
moderation strategies into new communities.

Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

Moderation Challenges in Voice-based Online Communities on Discord 55:3

2 Related Work
To situate our study, we begin by revisiting work on online community moderation, and voice
as an example of new technology adopted in online communities. These two threads of research
highlight how the unique characteristics of voice communication can not only bring new problems
to moderation, but also exacerbate existing problems. Finally we provide a brief review of research
on ephemerality, which we later show to be the source of the biggest challenge of moderating voice.

2.1 Moderation and Its Challenges
Effective regulation is one of the key factors that contributes to the success of an online community
[32]. Though regulation can occur at multiple levels, including Terms of Service (TOS), many online
communities have their own community-created rules that are enforced by volunteer moderators
[18]. Grimmelmann [22] lists four basic techniques that moderators can use to moderate their
communities: (1) excluding—to remove a member from the community; (2) pricing—to impose
monetary costs on participation; (3) organizing—to delete, edit, or annotate existing content; and (4)
norm-setting—to create desired, shared norms among community members. These techniques also
differ in terms of the ways they are used (e.g., automatically or manually, or proactively or reactively).
These techniques have become increasingly present in social computing and HCI scholarship,
focusing on how they have been successful in addressing various problems in communities (e.g.
[12, 27, 49]), ranging from using Reddit AutoModerator to automatically remove problematic
content [30], to setting positive examples to encourage similar behaviors in Twitch chat [48].
 While moderation is beneficial to communities, it is also challenging. Not only do moderators have
to do the often thankless job of continuously looking at traumatic content and making personally
uncomfortable decisions [17, 39, 45, 56], they also often have to resort to imperfect solutions to
problems in their communities. For example, while having clear, prominently displayed rules is
helpful for community members to learn the norms, it may convey the message that these rules are
often broken [32], or make community members feel stifled and constrained [30]. The lack of clarity
of the higher governing laws also made moderators’ work difficult [1]. Furthermore, the limited
availability of volunteer community moderators means that moderation is often delayed [34],
leaving problematic content in place, and community members’ efforts to circumvent moderation
[11, 20] makes timely moderation even harder. Recognizing this tension, prior research has called
for mutual understanding between community members and moderators [26].
 While there has been an emergence of machine-learning based automated moderation tools, it is
difficult to gather enough training data for rule violations when rule breakers try to hide themselves,
not to mention that these Automoderators may not be adaptive to new kinds of expression and
communication [21]. Even with the help of automated moderation tools, moderators still need to
make nuanced, case-by-case punishment decisions that automated tools are not capable of [49]. How
to scale this human labor is still an open and challenging question with today’s incredibly complex
online communities with millions of users of diverse backgrounds, purposes, and values [21]. Matias
[37] calls out online moderation work as civic labor to recognize this constant negotiation of the
meaning and boundary of moderation.
 These challenges are widespread in many online communities, and new ones often appear in the
context of new platforms and modes of interaction. For example, moderation practices on Twitch
had to adapt to fast-paced, real-time content in ways that would not be necessary on a platform like
Reddit [49, 56]. Our study continues to extend the current literature on moderation by examining
these challenges (and new ones) in the context of voice chat.

 Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

55:4 J. A. Jiang et al.

2.2 New Technologies and New Interactions
The introduction of new technologies to social interactions often results in unexpected challenges
in behavioral regulation–particularly as bad actors find new ways to break rules and push against
community norms. For example, in Julian Dibbell’s 1998 book My Tiny Life, he describes “a rape in
cyberspace” in which a user of the text-based community LambdaMOO used programmable avatars
to control and assault other users [16]. As a result, community moderators had to change their
rules and practices to deal with this unexpected use of the technology. New technology may result
in new structures, as people enact norms through their continued use of it [35, 42, 52]. Though
the technologies required to create programmable avatars were not new at the time they resulted
in bad behavior in LambdaMOO, it may have been the first time that community moderators had
to deal with how they might be used in that context. Similarly, voice communication technology
has been around for more than 100 years, but it is a newer consideration for online community
moderators.

2.2.1 Voice Communication. Group voice communication goes back as far as “party lines” on tele-
phone networks in the late 1800s. A party line was a shared telephone line for multiple households
that had to be manually connected by human operators. There was no expectation of privacy on a
party line because the operator as well as other people on the line could listen in at any time. As a
result, AT&T released guidelines around usage etiquette, encouraging people to cooperate. However,
these rules were “frequently broken,” and eavesdropping, gossiping, and pranking were common
problems [2, 13]. Though a very different context, we see analogs to the telephone company’s
struggle to moderate party lines in the struggles moderators shared in our study.
 Prior research about voice communication in HCI and CSCW has focused on affordances of voice,
often drawing comparison with video-based communication (e.g. [25, 40, 41]). This line of research
has revealed that video can increase the use of systems and results in better collaboration, but
still suffers from problems like cumbersome turn-taking coordination, lack of peripheral cues, and
difficult speaker determination (i.e., who “holds the mic”). The roots of these problems suggest that
they may be further exacerbated in voice-only communication due to the lack of visual cues, though
that social norms may be helpful for mitigating some of the problems. For example, in evaluating
real-time group voice communication as a social space, Ackerman et al. identified emerging norms,
including explicit announcement of someone new joining or of unpredictable noises [4]. Through
five case-studies, Wadley et al. [55] also showed that real-time voice was more susceptible to abuse,
a finding that our study also resonates.
 Recent research has examined voice communication in specific contexts. For example, Tang and
Carpendale [53] studied voice communication in a hospital, and identified the recurring problem
of ambient noise in voice communication. Another thread of research looked at voice-based
communities in rural India [43, 54]. Voice communication in these communities, however, is not
real-time but instead based on playback of recorded audio snippets. This important characteristic
not only made threaded conversations and actions such as rating and liking possible, but also
enabled moderation tools that exist in text-based communities, such as deleting, categorizing, and
ranking. More commonly, however, voice-based communication is real-time and ephemeral.

2.2.2 Ephemerality. Ephemerality is a distinctive feature of real-time voice communication, as well
as some contexts for text-based communication and online communities. Researchers have argued
that forgetting is a crucial part of human experience [38]. Ephemerality also facilitates spontaneous
interactions between users, encourages experimenting with different personas, and reduces con-
cerns about self-presentation [5, 47, 57]. Another line of research of anonymous communication
points out the ephemerality of identity and how it allows people to explore the full range of their

Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

Moderation Challenges in Voice-based Online Communities on Discord 55:5

identity but subject people to the consequence of de-anonymization [50], raising the question of
whether people can accurately estimate data persistence [28]. The answer to this question, however,
is that people tend to expect data persistence from platforms with default options being saving
rather than deleting [51], and come up with saving strategies to deal with content, meaning, and
context losses [7, 10]. While these prior research showed people tried to save content only for
personal consumption, persistent records are also critical to moderators’ work as evidence of rule
breaking. This study explores how gathering evidence becomes the biggest challenge in moderating
the ephemeral voice.

3 Research Site: Discord
Discord1 is a free cross-platform VoIP application that has over 200 million unique users as of
December 2018. Communities on Discord are called “servers,” a term we will use throughout this
paper to refer to these communities. Despite the term “server,” they are not self-hosted but instead
hosted centrally on Discord hardware. While originally designed for video gaming communities as
a third-party voice-chatting tool during gameplay, Discord servers now cover a wide range of topics
such as technology, art, and entertainment. Every user can create their own servers as they wish,
even simply as general chat rooms with no specific purpose. The size of Discord servers ranges
from small groups of friends with a handful of people, to massive communities with hundreds of
thousands of members.
 A server typically consists of separate virtual spaces called “channels,” usually with their own
purposes, such as announcements or topic-specific conversations. A channel can be either a text
channel or a voice channel, but not both. Users can also directly contact other users they are friends
or share servers with through direct messages with text, voice, and video capabilities. A screenshot
of the Discord interface is shown in Fig. 1.
 In voice channels, the only means of communication is real-time voice chat. Users can choose to
have their mic open all the time, or push a button to talk depending on their settings. Discord does
not provide ways to record or store voice chat, making them ephemeral. Users currently in a voice
channel will appear in the user list of the channel, and will disappear when they exit the channel.
A green circle around a user’s profile picture indicates the user is currently speaking. Users can
also mute themselves—make themselves not be heard—or deafen themselves—make themselves
not hear everyone else and not be heard. Some Discord servers also have a special type of voice
channel called “music queue,” where a music bot plays from a member-curated playlist, and all
other members are automatically muted.
 Server creators can create different “roles” with custom names that grant users different per-
missions in the server, through which moderators gain their permissions as well. This role system
allows for a hierarchy of moderation with lower-level moderators having less permissions, and
higher-level ones having more. Depending on permissions granted to a given role, moderators can
mute people, deafen people, or remove people from voice channels. Some moderators can also ban
people from their servers, who will not be able to rejoin unless they are “unbanned.”
 While the forms of punishment provided by Discord are permanent by default, third-party
applications called “bots” can be used to augment moderation by adding timers, making these
actions temporary. Bots like MEE62 , Dyno3 , and Tatsumaki4 are well-regarded and widely used by
over a million servers to automate existing Discord features such as sending welcome messages
and assigning roles. Besides improving existing moderator tools, many bots also provide additional
1 https://discordapp.com/
2 https://mee6.xyz/
3 https://dyno.gg/
4 https://tatsumaki.xyz/

 Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

55:6 J. A. Jiang et al.

Fig. 1. The Discord user interface. The far left sidebar lists all the Discord servers the user is a member of.
 The next side bar lists the text and voice channels of the Discord server the user is currently viewing. The
middle area is for the scrolling text chat, and the right side bar lists the total users, categorized by their “role.”

functionalities for moderators, such as issuing people warnings that are permanently recorded in a
moderator-only channel, and automatically removing content in text channels based on keywords
or regular expressions. However, to the best of our knowledge, there are currently no bots with
voice moderation capabilities.

4 Method
To understand moderators’ experiences in moderating voice-based communities, we conducted in-
depth, semi-structured interviews with moderators of Discord servers. Participants were recruited as
part of a larger collaborative, IRB-approved project to investigate moderation in online communities.
For this study we analyzed 25 interviews with moderators who identified as having experience in
moderating Discord voice channels. We recruited participants by reaching out to moderators of
open Discord servers. We also asked them to send the call for participation to other moderators,
resulting in a snowball sample. The first two authors conducted the interviews. The 25 participants
came from 16 different Discord servers, with between 1 and 3 participants from each server. While
the majority of the servers that we examined are large ones with more than one thousand members
and may not be representative of smaller groups, we believe this over-representation is reasonable
as formal moderation is less-needed in smaller communities [29, 49]. Our moderator participants
provided us with a diversity of perspectives both across and within communities. Each participant
was compensated US $20 for their time.
 Interviews ranged in length from 42 to 97 minutes, all of which were conducted over Discord voice
chat. Participants’ ages ranged from 18 to 39 ( = 24, = 5.43). Details about the participants,
including age, gender, country of residence, and type and member count of the servers they moderate
are presented in Table 1.

Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

Moderation Challenges in Voice-based Online Communities on Discord 55:7

 Table 1. Participant details. Numbers of server members are as of March 8, 2019.

 Participant ID Age Gender Country Server Type # Members
 P01 18 M Croatia Social Chatroom 164,257
 P02 19 M US
 Streamer 117,742
 P03 19 M Russia
 P04 21 M US Tech Support 233,762
 P05 21 M India
 Anime 130,924
 P06 20 M US
 P07 18 M UK
 Social Chatroom 57,319
 P08 20 F Malaysia
 P09 22 M US NSFW 23,186
 P10 23 M UK
 Gaming 29,165
 P11 39 M UK
 P12 23 F US Fandom 150
 P13 24 M Australia NSFW 55,239
 P14 19 M US Social Chatroom 77,512
 P15 26 M US
 Gaming 55,251
 P16 24 M US
 P17 37 M US
 Fiction Writing 1,137
 P18 32 F US
 P19 26 F US Gaming 3,246
 P20 24 F Netherlands
 P21 27 M US Gaming 24,542
 P22 22 F US
 P23 23 M Netherlands Gaming 171,608
 P24 24 F UK
 Gaming 63,001
 P25 29 M US

 During the interviews, we asked participants to tell us specific stories about moderating voice
channels, with follow up questions about how they found out about rule breaking, what specific
actions they took, and what impact the incident had on the moderation team as well as the
community. We also asked them to consider hypothetical scenarios, such as what participants
would do if the rule breakers tried to evade punishment. Participants detailed a variety of moderation
experiences that ranged in scale and in complexity. We also asked participants about the challenges
of moderating voice channels, the behind-the-scene deliberations of their moderator teams, and
their feelings toward decisions they had made or situations they had encountered. Prior to analysis,
all interviews were transcribed, anonymized, and assigned the participant IDs presented here.
 We performed a thematic analysis of the interview transcripts [9]. The first author initially
engaged in one round of independent coding, using an inductive open coding schema. All authors
then discussed preliminary emerging code groups such as “catch in the act,” or “enter the voice

 Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

55:8 J. A. Jiang et al.

channel to confirm.” Two more rounds of iterative coding helped us combined similar groups to
create higher order categories such as “moderation challenges.” The first author used these categories
to produce a set of descriptive theme memos [46] that described each category with grounding in
the interview data. All authors discussed the memos regularly to reveal the relationships between
the categories and finally clarified the themes, which resulted in the three main findings we discuss
below.

5 Findings
In describing the findings of this study, we start by characterizing new types of rules and new ways
to break these rules in voice channels, then compare them to common types of rule violations in text
communication. We then discuss the actions that moderators take to address these rule violations.
Finally, we address the biggest challenge of rule enforcement in voice—acquiring evidence—by
discussing moderators’ strategies to gather evidence and how they often fail.

5.1 Rules in Voice and How People Break Them
Formal rules on Discord exist at the platform level in the form of Terms of Service and Community
Guidelines, as well as at a community level in the form of custom rules set by the individual Discord
servers. All the servers in our study had at least some explicit rules that were listed in specialized
text channels, as well as implicit rules that were not written down but were nevertheless enforced
by moderators. Though there were likely also emergent social norms in these communities, and
rules that may have started out as norms, we spoke to moderators about the rules that they actively
enforced, whether explicit or implicit, as opposed to norms enforced by the community itself. While
there were many rules in the servers we examined, here we only focus on those with elements
unique to voice.

5.1.1 Explicit Rules. Servers that we discussed with moderators had different explicit rules that
governed both text and voice channels, such as “no advertising” or “English only,” but all 16 of them
had a rule against slurs and hate speech. We choose to take a deep dive on the rule of slurs and hate
speech because it is the rule that most participants talked to us about, and presented challenges
unique to voice.
 Slurs and hate speech can make a community an unwelcoming and unsafe space for its mem-
bers, and therefore many communities have rules against them [18]. Just like in many text-based
communities, slurs and hate speech are explicitly prohibited in voice channels, and are a major
problem that moderators have to face. All participants told us that racial and homophobic slurs
existed widely in their servers, both text and voice channels. In P08’s server, racial slurs in voice
channels faced an even harsher punishment than in text channels:
 Racial slurs in the [text] chat and VC [voice chat] are different. If you say it in the [text]
 chat, you get a four-hour mute depending on the severity, and in the VC, you get an instant
 ban because it’s more ... you know, saying it, rather than typing it, is much worse. (P08)
 Racial slurs can be more offensive when spoken in smaller groups. Voice channels usually have
5 to 25 people participating at a time, which is much less than in text channels that typically have
hundreds of active members. A potential consequence of the limited number of participants is that
slurs in voice channels may feel more targeted and personal.
 While slurs were not allowed in any of the servers, how moderators determined the threshold
for what counted as a slur varied. For example, P03 treated the slur “n---er”and all of its intonations
with a heavy hand:

Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

Moderation Challenges in Voice-based Online Communities on Discord 55:9

 Like if you were saying, the “N” and then “word,” then it’s fine because it’s not saying
 the slur itself. Any workaround ... is not allowed. “N---a” or “n---a”—that’s not allowed.
 Because you’re still saying the slurs. Just rephrasing it. (P03)
 Many moderators were cognizant of the different intonations a slur can have, and still decided
to uniformly ban them—the only exception was the indirect reference “n-word.” At the same time,
while P06 agreed that different varieties of racial slurs were still racial slurs, he also took context
and intention into account in his moderation, and the intonations did matter in his decisions:
 I think there’s a difference between saying “What’s good my n---a” and “I think all n---ers
 should go back to Africa.” There’s a pretty clear difference. So in that sense, you know,
 they’re both racial slurs technically. And so by our rules ... the punishment would be the
 same for both. But you know, again, it’s kind of like a case-by-case thing. (P06)
 While the intonations can still be expressed in text, P06’s quote suggests that voice introduces
more nuances to moderation, and that what technically counts as racial slurs by explicit rules may
still receive different treatment. In text-based communities, moderation of content and intonations
can be automated by using a list of keywords or regular expressions (e.g., [11]). But in voice
communication where moderation cannot be as easily automated, and moderators have to hear
everything for themselves, their personal judgments play a more important role. Having to moderate
in a case-by-case manner also means more work for moderators.
5.1.2 Implicit Rules. While slurs and hate speech were explicitly against the rules in the Discord
servers we examined, we also heard about behaviors that were not written in the rules, but that
moderators still discouraged or prohibited. While moderators technically had the option to detail
these in their explicit rules, these behaviors were complex, nuanced, and difficult to articulate.
Below, we focus on three main types of these behaviors that are unique to voice, and describe them
as implicit rules in the servers: disruptive noise, music queue disruptions, and raids.
 5.1.2.1 Disruptive Noise. Disruptive noise involves intentionally creating a loud or obnoxious
sound in voice channels to irritate other people and disrupt conversations. According to many
moderators, disruptive noise is a common rule violation in voice channels. One moderator, P14,
said that their typical day involves “muting at least one person” in response to this kind of behavior.
Disruptive noise points to several implicit rules that are not important in text-based communities,
but stand out in voice. One of these rules is that one should not speak too loudly in a group
conversation:
 I’ve had to step in because someone’s told me “Oh there’s a kid literally screaming down
 in the #underbelly” ... So I’ll hop in, and of course the kid will be screaming and he’ll be
 muted. (P16)
 The rule of not speaking too loudly shows a key difference between voice and text communication:
voice is a limited-capacity communication channel. While typing in all caps does not affect other
members’ ability to type and be seen, speaking in high volume in a group voice channel takes up all
capacity in the channel, effectively silencing others. Even though text spamming—posting lots of
content repeatedly and quickly—also takes up channel capacity in a similar way, it is preventable by
limiting the number of messages one can post in a given period of time (e.g., “slow mode” in Discord).
Loud screaming, on the other hand, is not actively preventable on Discord unless a moderator steps
in. Prior work has shown that people are already struggling with how to appropriately interject in
group conversations [25] with their assumed sequential nature [15]. The rule violation here is even
more severe because it completely ignores turn-taking in conversations and forces the conversation
to be about one person.

 Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

55:10 J. A. Jiang et al.

 However, volume by itself was not the golden rule of determining whether someone is creating
noise. As P14 suggested, hardware conditions also had an impact on someone’s speaking volume:
 [Disruptive noise is] typically anything that would be considered “too loud.” However, if
 someone has a sensitive mic, we typically let them know of this and make them aware
 how loud it is before taking action. (P14)
 P14 further told us how he differentiated between intentional noise making and simply having a
sensitive mic:
 When someone joins a VC [voice channel] for the sole purpose of “ear raping”5 , they will
 typically do so immediately; someone who joins a VC to use it for its intended purpose will
 typically start off with a greeting or some sort of intro. I try to be aware that not everyone
 is capable of purchasing the best microphones, so I try to take that into consideration. (P14)
 While it may be possible to develop an automated program that detects disruptive noise by
decibel value, P14’s quotes show that in addition to presented behaviors, intention also matters. If
intention is important for determining rule violations, automated tools clearly have limitations.
Furthermore, the fact that the way someone initially joins a voice channel is important suggests a
challenge in moderation: A moderator has to be present when someone joins, which is not a scalable
solution when voice channels are always open and a moderator is always required. This constant
presence is also not a reasonable request for volunteer moderators who are only contributing
during their free time.
 In addition to the appropriateness of volume, disruptive noise as a type of rule violation also
points to the appropriateness of content in group conversation:
 There is one time I had my Discord on speaker and I was just talking to a group of friends.
 ... [Some random people] joined and they started playing loud porn. So my brother was in
 the house ... and he heard the porn blasting out of my speakers and he was like, “Yo dude,
 why are you listening to that at full blast on speaker?” (P06)
 People tend to dislike unexpected auto-play audio on computers as it can cause physical dis-
comfort and can be socially awkward in public [4, 14]. The inappropriateness of the content here
only exacerbated the situation. Furthermore, people not directly part of the voice channel were
also affected, which demonstrates how moderation work can potentially have a negative impact on
moderators’ lives outside the servers they moderate. P06 told us that the experience was “a little
bit awkward,” but it is likely that the same situation can happen in other contexts, such as in a
workplace, with more serious consequences.
 We also heard a similar story of disruptive noise where the sound itself was not irritating, but
rather the way the sound was played:
 We’d have this weird phenomenon where there was a handful of people who would join
 voice for 5 seconds, leave for 2 minutes, then come back and join voice for 5 seconds, leave.
 So while you’re playing all you would hear was “boop” “do-doop” [Discord notification
 sounds]—people just joining and leaving rapidly—and it was just so infuriating. (P10)
 Just like in P06’s example above, the unexpectedness also stands out in this example—people
would not expect Discord features to be used in annoying ways. Furthermore, while it is possible
to turn off these notification sounds, it is not feasible to do so in the middle of a game. There is also
no way to ping a moderator without disengaging with the ongoing game, as in the case in P10’s
5 Though we intentionally excluded it from all but this quote, we would like to note that a number of our participants used
the term “ear raping” to refer to creating disruptive noises. We hope that future work examines the use of the term in more
depth, including the motivations behind its use and in what ways it may be harmful to individuals or communities on the
site.

Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

Moderation Challenges in Voice-based Online Communities on Discord 55:11

quote. This example also points to a difference between text and voice—in text channels, constant
interruptions are accepted and somewhat expected. But in voice channels, the flow is much more
important and even the slightest interruption can be disruptive.

 5.1.2.2 Music Queue Disruption. Across our interviews, we heard stories of rule violations not
only in conversational voice channels, but also in music queues—voice channels that automatically
play from crowdsourced playlists. In music queues, members are automatically muted, but they
can append new music to the shared playlist, and skip the music that is currently playing if the
channel setting allows.
 A rule that music queues shared with conversational voice channels was not to be too loud:
 Literally the most recent thing that happened. ... Someone put [something] in the music
 queue and it was for like two hours of just extremely loud music. (P04)
 P04’s quote shows that there was another type of rule violation other than being too loud: the
music was also long. According to P04, checking whether people put “hour-long shitposting tracks
that nobody wants to listen to” in the music queue was his everyday job. Because music queues are
necessarily sequential, a turn-taking rule becomes important: people agree to occupy the channel
completely for an appropriate, limited amount of time. Compared to disruptive noise, the problem
with playing long music is not that it takes up the whole channel—it is in fact expected—but is that
it essentially rids other members of their turns by taking up the channel for a long time.
 Another form of rule violation that stops other people from participating is to skip their music:
 I actually got mad at the user. ... What they were doing was they were constantly taking
 our jukebox and turning off other people’s stuff. Like someone would play something and
 then they would skip it. (P21)
 This example, together with the previous example, emphasizes the importance of turn-taking
in music queues. While turn-taking is not a problem in text because of its threading structure, it
becomes a major problem in music queues that resemble town hall meetings, where everyone gets
an opportunity to hold the mic—playing long music is like someone who will not stop talking, and
skipping music is like forcing the person with the mic to stop talking.

 5.1.2.3 Raids. In addition to rule violations by individuals, we also heard stories of organized
rule violations that moderators called “raids.” Raids are organized voice channel disruptions that
involve multiple users that can be human or bots. P13, for example, experienced a raid that was
similar to P06’s story, but on a larger scale:
 There was one a few months ago where they were spamming porn over their mics and they
 all had profile pictures of the same girl in a pornographic pose. And there were maybe like
 15 of them in the same voice chat. (P02)
 While an individual playing pornographic audio is irritating by itself, one can only expect 15
people doing so would cause even more discomfort. While raiding violates the similar rules in voice
that we mentioned above, it is considerably more difficult for a moderator to manage. In the case
of an individual, a moderator only needs to take action on that person, but in the case of a raid, a
moderator needs to act on all the people involved. Across our interviews, we heard stories of raids
that involved up to thousands of bot accounts, but moderators could only manage the accounts
one by one—there is currently no way to manage multiple accounts all at once. This restriction
means that not only do the moderators have to take on a significant amount of work managing
raids, but also there is no way to prevent the other raiders from evading once they see one of them
is punished.

 Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

55:12 J. A. Jiang et al.

 Fortunately, while managing raids could be difficult, recognizing raids was relatively easy. As
P02’s quote suggested, the raiders all had similar profile pictures, which often became a clear signal
for raid detection:
 [W]hen I see multiple people with the same or similar names rapidly join a VC, that’s a
 warning sign for me. (P14)
 Moderators told us that when they saw people with similar profile pictures or names, they
would join the voice channel to confirm if they were part of a raid. However, we also heard from
some moderators that they saw these cues as definitive signals of raids, and punished these people
without confirming. Moderators told us that there were no unfair punishments because they
believed if there had been any, these people would have appealed. This moderation approach can
be potentially problematic in cases where members use similar usernames as part of a community
in-joke, which we also heard in the interviews. While moderators’ tolerance of false positives—
incorrectly identifying someone as a raider—may be a reasonable attempt to reduce their workload,
it also suggests that they could have unknowingly driven these members away for good.

5.2 Moderation Practices
To punish the rule breakers, moderators used tools provided by Discord, including muting, deafening,
and banning for more serious offenses, which third-party bots enhanced by setting a timer on
them, making these punishment temporary. While we found that the correspondence between rule
violations and types of punishment was highly contextual to individual servers, there were still
some common moderation approaches that the moderators took. Specifically, moderators tried to
warn first before taking more severe actions, but sometimes they also took these actions merely
based on hearsay or first impressions. Even though these signals were unreliable, moderators
did not come up with specific rules that described acceptable and unacceptable behaviors, but
developed catch-all rules that dictated their authorities instead.
5.2.1 Warn Before Punishment. Across our interviews, a common approach to moderate voice
channels that we heard from 23 of 25 moderators was to warn first before giving out real punishment:
 We would unmute our mics, start talking casually to people and then we would just figure
 out if they are breaking the rules or not, warn them verbally in the voice channel because
 we do roll with things in voice channel. ... We’d always do verbal warning before we do
 any other actions. (P07)
 P07’s quote points to a key difference between voice moderation and text moderation. In the case
of text, moderators would not have to “figure out” if someone broke the rules—all the conversations
are recorded and visible to moderators. However, because Discord does not record voice channels,
there is no way for moderators to unambiguously determine rule violations. “Rolling with things”
allowed moderators to learn the contexts of alleged rule violations while not upsetting community
members, and giving warning first let moderators still enforce rules to some extent but not have to
risk wrongly punishing someone.
 According to the moderators, warning was not only a way to conservatively enforce rules, but
also a lesson for the community:
 We have this weird thing where our moderation is kind of public in that, if we warn
 someone not to do that, we try to always do that publicly so that other people can say,
 “Hey, okay, that’s an example of what not to do.” (P19)
 The public moderation that P19 mentioned suggests community members learn rules through
examples, which prior research shows can be effective in discouraging undesired behaviors [48].

Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

Moderation Challenges in Voice-based Online Communities on Discord 55:13

While example-setting certainly has been successful in moderating text-based communities, ex-
plicitly showing what the rules are may be more important in voice-based communities because
the rules might be ambiguous or unfamiliar, especially to newcomers who are more familiar with
text-based communities.
5.2.2 Punishment Based on Hearsay and First Impressions. While in most cases moderators tried to
give warnings first, they sometimes also punished rule breakers directly. For example, P11 told us
about a time the moderation team banned a member only based on a member report, without extra
deliberation:
 Someone complained that a user was harassing them in voice chat and just shouting
 profanities and racism down the mic, so we just banned the user and we didn’t hear
 anything, and they didn’t appeal it. ... We kind of took it at face value. It’s very hard to
 get evidence from a voice chat unless you’re recording. (P11)
 Here, P11’s moderation team assumed the punishment was fair only because the person punished
did not push back, which may be an acceptable solution without adequate evidence. However,
punishment based on hearsay does risk wrongly punishing someone in the case of false reporting.
In such a case, the person punished by mistake possibly would not have appealed anyway when
they were frustrated by undeserved punishment and left the community [33].
 Moderators sometimes also took actions based on their own understanding of the person involved.
P21, for example, told us that “first impressions matter, especially in voice chat,” to the extent that
they justified the most severe form of punishment:
 So if a user has just joined ... and immediately shows this kind of behavior, such as the
 derogatory terms, trolling, sexist behavior ... they’ll just get banned. We won’t warn them.
 We’ll ban you because you’ve made it quite apparent that you had no interest in reading
 the rules and you’re not all that great of a person, to say the least. (P21)
 P21’s quote shows a stark contrast with the “warning first” approach of moderation, which
he also took in his community. However, the importance of first impressions in voice suggests
that a person’s intention and character may be more salient than in text, and they can be used in
moderation decisions.
5.2.3 Catch-all Rules. With moderators sometimes having to use unreliable cues like first impres-
sions in moderation, one question arises: Couldn’t they develop rules that unambiguously describe
what behaviors were acceptable or not? Moderators gave their answers to this question:
 There are times where you can’t catch everything with a rule set. So I think that’s more
 the reason why we don’t have an on paper set of rules because we are such a large server.
 We would have to have a million different rules to help cover everything. (P04)
 Across the Discord servers we examined, only two had guidelines about acceptable behaviors
specific to voice channels. P04’s quote shows that the lack of guidelines comes from the practical
standpoint that it is simply impossible to list every single type of behavior, which suggests that the
variety of (mis)behaviors possible in voice is much more diverse than in text. However, this is not
only a problem of quantity—the level of nuance involved in the rule violations themselves shows
that the line between acceptable and unacceptable behaviors is also difficult to articulate.
 The solution that servers used to solve this problem, including the two that had guidelines in
voice channels, was to create catch-all rules that were different varieties of the same core idea:
Moderators have the utmost authority.
 [If someone does something unacceptable] then they are well within their rights to turn
 around and say, “Well that isn’t in the rules.” And then it’s just a nice get-out clause for

 Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

55:14 J. A. Jiang et al.

 the moderator to be able to turn around and say, “Well look, whatever a moderator says
 goes.” (P11)
 This umbrella rule did not only free the moderators from having to create new rules, but also
reduced the potential work of arguing with members. As moderators had greater authority, they also
took greater responsibility in making the right decisions. But as we heard from the moderators, the
lack of evidence in voice became an obstacle. In the next section, we describe the ways moderator
gathered evidence of rule violations and their unique challenges.

5.3 Acquiring Evidence
Moderators told us about their various practices and actions upon rule violations, but a prerequisite
of any action is that moderators had to make sure that a rule was violated. Acquiring such evidence
in ephemeral voice channels, however, was a major challenge:
 Voice channels just basically can’t be moderated. ... The thing is in voice, there’s no record
 of it. So unless you actually heard it yourself, there’s no way to know if they really said it.
 (P19)
 Gathering evidence was such a challenge that moderating voice channels was almost impossible
to P19, and this difficulty points directly to a fundamental difference between voice and text. In
text, it is common knowledge that every conversation is automatically recorded. Moderators would
not even have to consider acquiring evidence because everything is persistent. Even for ephemeral
text chat applications like Yik Yak, where community moderation mechanisms such as upvoting
and downvoting are possible [47], the text would still have to be persistent for a short period of
time. However, this common assumption about text-based moderation breaks down completely in
voice because of the ephemerality of real-time voice conversation.
 Moderators came up with different strategies and workarounds to tackle the problem of obtaining
evidence in an ephemeral environment. In the rest of this section, we describe three main types
of strategies that we heard from moderators: (1) Entering voice channels when the rule violation
was happening; (2) asking witnesses to confirm someone was breaking the rules; and (3) recording
voice channels.
5.3.1 Entering Voice Channels to Confirm Rule Breaking. As P19 said, the most reliable way for
moderators to acquire evidence was to hear it for themselves. However, constant presence in voice
channels is not a reasonable request for volunteer moderators. Therefore, moderators took one
step back—entering voice channels when they realized someone might be breaking the rules:
 When we do get a report, then if someone is available and it’s going on right then and
 there, then we’ll hop on and listen and see what’s going on and see if we can determine
 any rule violations right there off the bat. (P25)
 While entering voice channels may be a solution, P25’s quote does point out a important
requirement: a moderator has to be online at the time of rule violation, and the rule violation has
to be still ongoing when the moderator joins. This requirement is difficult to fulfill, considering
the time a member would take to report to a moderator, the time the moderator takes to see the
report, and the time the moderator takes to join the voice channel. This requirement also means
that, for any violations that are instant and do not extend over a period of time, moderating with
this method is nearly impossible.
 Even if a moderator was present at the right time, it was still not guaranteed that the moderator
could identify the rule breaker:
 There’s also the problem of telling who’s talking, because ... if everyone’s talking at once
 and some big fight is happening, it’s kind of hard to tell who’s talking. (P21)

Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.

Moderation Challenges in Voice-based Online Communities on Discord 55:15

 Discord has no platform limit on the number of people a voice channel can have, and many mod-
erators had moderated large voice channels with as many as 25 people. Voice channels, unlike text
channels with threading structures, is inherently limited in information capacity. Therefore, it can
be difficult to tell who is saying what when many people are talking simultaneously. Furthermore,
there was also another problem with multiple people in the same voice channel—there was no way
to stop people from evading:
 When they start seeing people get banned, they leave the voice chat so they don’t get
 banned themselves and then it makes it harder for us to find them. (P02)
 Tracking down rule breakers, according to P20, was often a “wild goose chase, only on steroids.”
The need to track down rule breakers speaks to the importance of not only acquiring evidence, but
also the information contained in the evidence, which is also different between in text and in voice.
In text, the text itself contains evidence in the form of rule-breaking content (e.g., racial slurs), but
also the metadata—who the poster was, when it was posted, etc. When a moderator need to punish
a person, this information connects the problematic content to the person, allowing the moderator
to act without the person and the content being present at the same time. In real-time voice where
content is ephemeral, however, this metadata does not exist. Even though the moderator has the
evidence—they heard the content—there is no persistent record that associates the content to the
rule breaker. Furthermore, identities in voice channels are also ephemeral: once the person leaves
the voice channel, their profile picture disappears, and the information that connects the voice
to the person no longer exists. Without this metadata, in text the moderator can at least delete
the content, but in voice, the content is ephemeral to start with, and the moderator can neither
moderate the content nor punish the person.
5.3.2 Relying on Witness Reports. With joining voice channels being ineffective in many cases,
some moderators turned to witnesses’ testimony as evidence:
 If the violation has already passed and people have split up, then we’ll get a list of the
 users that were in that voice channel at the time and we’ll send a message to them and
 just ask them, “Hey, you heard that there was something going on with this voice channel,
 can you give us more information about this?” (P25)
 Moderators hoped to use witnesses’ memories as a way to overcome the ephemerality of voice,
but note that P25 here contacted a list of members instead of a single member. Without concrete
evidence, the quality of witness reports became important to separate legitimate reports from
hearsays and rumors, and one proxy for quality was the number of witnesses:
 We need witnesses like at least maybe two other people ... that were there, that can confirm
 like, “Yes, this person said that.” (P02)
 P02’s quote suggests a simple rule of thumb: the more witnesses there were, the more credible
the report was. While moderators spoke of this as a generally successful approach, there were
nevertheless problems with it. First, multiple witnesses would work only if their stories were
consistent, but when they were not, moderators were likely to fall into the rabbit hole of seeking
more witnesses and more stories. This could lead to the moderation workload no longer being
comparable to the severity of the violation anymore. Second, even if the stories were consistent,
moderators had no way to make sure the stories were true:
 You just have to be careful with the report, because some people can do false reporting, to
 do whatever to the other user. (P24)
 P24’s quote reveals an interesting possible scenario, where one member falsely accuses another
member due to some personal vendetta. Requiring multiple witnesses may be able to mitigate this
problem between dyads, but could facilitate a planned brigade against a community member, and in

 Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 55. Publication date: November 2019.